Hip tensor permute unit test (#1068)

* adding files for F32 example

* adding functioning implementation with scalar multiplication and unary operator support

* added fp 16 type check in unary square

* updating scalar multiplication as an operator

* functioning version with scalar operator

* changing strides for col major

* updated column major implementation

* working column major implementation

* cleaned up comments, rearranged/renamed files

* small edits to 3d transpose profiler

* adding test/profiler/instance files for hipTensor permute unit test

* added more test instances

* cleaned up errors, randomized input tensor, added more instances

* turned off time printouts

* removed conflicting transpose profiler

* rearranged some files

example/44_elementwise_permute/elementwise_permute_4D_fp16_col.cpp

 #include <iostream>
 #include <cstdlib>
+#include <random>
 
 #include "ck/ck.hpp"
 #include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
@@ -48,10 +49,8 @@ void host_elementwise4D(HostTensorB& B_nhwc,
                 for(std::size_t n = 0; n < N; ++n)
                 {
                     ADataType tmp_val;
-                    // auto a_val = A_nchw(n, c, h, w);
                     auto a_val = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
                     functor_b(tmp_val, a_val);
-                    // functor_a(B_nhwc(n, h, w, c), scale * tmp_val);
                     functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
                               scale * tmp_val);
                 }
@@ -62,12 +61,14 @@ int main()
     bool do_verification = true;
     bool time_kernel     = true;
 
-    std::vector<std::size_t> nchw = {4, 2, 1, 8};
-    std::vector<std::size_t> nhwc = {4, 1, 8, 2};
+    std::vector<std::size_t> nchw = {16, 8, 32, 64};
+    std::vector<std::size_t> nhwc = {16, 32, 64, 8};
     Tensor<ADataType> a(nchw);
     Tensor<BDataType> b(nhwc);
     float scale = 1.f;
     auto i      = 0;
+    std::mt19937 gen(11939);
+    std::uniform_int_distribution<int> dis(0, 1);
     for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
         for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
             for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
@@ -75,7 +76,7 @@ int main()
                 {
                     a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
                             (h * nchw[3]) + w] = i;
-                    i++;
+                    i                          = dis(gen);
                 }
 
     DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());

example/44_elementwise_permute/elementwise_permute_4D_fp32_col.cpp

@@ -67,6 +67,8 @@ int main()
 
     float scale = 1.f;
     auto i      = 0;
+    std::mt19937 gen(11939);
+    std::uniform_int_distribution<int> dis(0, 1);
     for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
         for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
             for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
@@ -74,7 +76,7 @@ int main()
                 {
                     a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
                             (h * nchw[3]) + w] = i;
-                    i++;
+                    i                          = dis(gen);
                 }
 
     DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());

library/include/ck/library/tensor_operation_instance/gpu/permute_scale.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_permute_scale_f16_instances(
+    std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F16>,
+                                                  ck::Tuple<F16>,
+                                                  PassThrough,
+                                                  element_wise::UnarySquare,
+                                                  Scale,
+                                                  4>>>&);
+
+void add_device_permute_scale_f32_instances(
+    std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>,
+                                                  ck::Tuple<F32>,
+                                                  PassThrough,
+                                                  element_wise::UnarySquare,
+                                                  Scale,
+                                                  4>>>&);
+
+template <typename InDataTypeTuple,
+          typename OutDataTypeTuple,
+          typename ElementwiseOperation,
+          typename UnaryOperation,
+          typename Scale,
+          index_t NumDim>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::DeviceElementwise<InDataTypeTuple,
+                                                    OutDataTypeTuple,
+                                                    ElementwiseOperation,
+                                                    UnaryOperation,
+                                                    Scale,
+                                                    NumDim>>
+{
+    using DeviceOp = DeviceElementwise<InDataTypeTuple,
+                                       OutDataTypeTuple,
+                                       ElementwiseOperation,
+                                       UnaryOperation,
+                                       Scale,
+                                       NumDim>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InDataTypeTuple, ck::Tuple<F32>> &&
+                     is_same_v<OutDataTypeTuple, ck::Tuple<F32>>)
+        {
+            add_device_permute_scale_f32_instances(op_ptrs);
+        }
+        else if constexpr(is_same_v<InDataTypeTuple, ck::Tuple<F16>> &&
+                          is_same_v<OutDataTypeTuple, ck::Tuple<F16>>)
+        {
+            add_device_permute_scale_f16_instances(op_ptrs);
+        }
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/permute_scale/CMakeLists.txt

+add_instance_library(device_permute_scale_instance 
+	device_permute_scale_instances.cpp)

library/src/tensor_operation_instance/gpu/permute_scale/device_permute_scale_instances.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+#include "ck/utility/data_type.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Pass    = ck::tensor_operation::element_wise::PassThrough;
+using UnaryOp = ck::tensor_operation::element_wise::UnarySquare;
+using Scale   = ck::tensor_operation::element_wise::Scale;
+
+// clang-format off
+using device_permute_scale_f16_instances =
+    std::tuple <
+        DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 1, ck::Sequence<1>, ck::Sequence<1>>,
+        DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 8, ck::Sequence<1>, ck::Sequence<1>>,
+        DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
+        DeviceElementwiseImpl<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4, 2, ck::Sequence<1>, ck::Sequence<1>>
+    >;
+
+using device_permute_scale_f32_instances = std::tuple<
+         DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 1, ck::Sequence<1>, ck::Sequence<1>>,
+         DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 8, ck::Sequence<1>, ck::Sequence<1>>,
+         DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 4, ck::Sequence<1>, ck::Sequence<1>>,
+         DeviceElementwiseImpl<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4, 2, ck::Sequence<1>, ck::Sequence<1>>
+    >;
+// clang-format on
+
+void add_device_permute_scale_f16_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, Pass, UnaryOp, Scale, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f16_instances{});
+}
+
+void add_device_permute_scale_f32_instances(
+    std::vector<std::unique_ptr<
+        DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, Pass, UnaryOp, Scale, 4>>>& instances)
+{
+    add_device_operation_instances(instances, device_permute_scale_f32_instances{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/transpose/device_transpose_instances_3d.cpp

@@ -19,22 +19,14 @@ void add_device_transpose_f16_instances(
     std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F16>, ck::Tuple<F16>, PassThrough, 5>>>&
         instances)
 {
-#ifdef CK_ENABLE_FP16
     add_device_operation_instances(instances, device_transpose_f16_instances{});
-#else
-    ignore = instances;
-#endif
 }
 
 void add_device_transpose_f32_instances(
     std::vector<std::unique_ptr<DeviceElementwise<ck::Tuple<F32>, ck::Tuple<F32>, PassThrough, 5>>>&
         instances)
 {
-#ifdef CK_ENABLE_FP32
     add_device_operation_instances(instances, device_transpose_f32_instances{});
-#else
-    ignore = instances;
-#endif
 }
 
 } // namespace instance

test/CMakeLists.txt

@@ -150,6 +150,7 @@ add_subdirectory(batched_gemm_multi_d)
 add_subdirectory(grouped_convnd_bwd_data)
 add_subdirectory(conv_tensor_rearrange)
 add_subdirectory(transpose)
+add_subdirectory(permute_scale)
 add_subdirectory(wrapper)
 if(GPU_TARGETS MATCHES "gfx11")
     add_subdirectory(wmma_op)

test/permute_scale/CMakeLists.txt

+add_custom_target(test_permute)
+add_gtest_executable(test_permute_scale test_permute_scale.cpp)
+if(result EQUAL 0)
+  target_link_libraries(test_permute_scale PRIVATE utility device_permute_scale_instance)
+  add_dependencies(test_permute test_permute_scale)
+endif()

test/permute_scale/test_permute_scale.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "gtest/gtest.h"
+#include "test_permute_scale_impl.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using ck::index_t;
+
+template <typename Tuple>
+class TestPermute : public ::testing::Test
+{
+    protected:
+    using ADataType = std::tuple_element_t<0, Tuple>;
+    using BDataType = std::tuple_element_t<1, Tuple>;
+
+    void Run()
+    {
+        std::vector<std::vector<ck::index_t>> lengths = {
+            {4, 2, 1, 8}, {1, 1, 1, 1}, {16, 8, 32, 64}, {32, 64, 128, 128}};
+
+        for(auto length : lengths)
+        {
+            bool success =
+                ck::test_permute_scale_impl<ADataType, BDataType, 4>(true, 2, false, false, length);
+            EXPECT_TRUE(success);
+        }
+    }
+};
+
+using KernelTypes = ::testing::Types<std::tuple<F16, F16>, std::tuple<F32, F32>>;
+
+TYPED_TEST_SUITE(TestPermute, KernelTypes);
+TYPED_TEST(TestPermute, Test_FP16) { this->Run(); }
+TYPED_TEST(TestPermute, Test_FP32) { this->Run(); }

test/permute_scale/test_permute_scale_impl.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iomanip>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_elementwise_scale.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_scale_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/permute_scale.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+namespace ck {
+template <typename HostTensorA, typename HostTensorB, typename FunctorA, typename FunctorB>
+void host_elementwise4D(HostTensorB& B_nhwc,
+                        const HostTensorA& A_nchw,
+                        FunctorA functor_a,
+                        FunctorB functor_b,
+                        float scale)
+{
+    std::size_t N = A_nchw.mDesc.GetLengths()[0];
+    std::size_t C = A_nchw.mDesc.GetLengths()[1];
+    std::size_t H = A_nchw.mDesc.GetLengths()[2];
+    std::size_t W = A_nchw.mDesc.GetLengths()[3];
+    for(std::size_t w = 0; w < W; ++w)
+        for(std::size_t h = 0; h < H; ++h)
+            for(std::size_t c = 0; c < C; ++c)
+                for(std::size_t n = 0; n < N; ++n)
+                {
+                    using tmp_type   = ck::remove_reference_t<decltype(B_nhwc(0, 0))>;
+                    tmp_type tmp_val = 0;
+                    auto a_val       = A_nchw.mData[(n) + (c * N) + (h * C * N) + (w * H * C * N)];
+                    functor_b(tmp_val, a_val);
+                    functor_a(B_nhwc.mData[(n) + (c * W * H * N) + (h * N) + (w * H * N)],
+                              scale * tmp_val);
+                }
+}
+
+template <typename ADataType, typename BDataType, index_t NumDim>
+bool test_permute_scale_impl(int do_verification,
+                             int init_method,
+                             bool do_log,
+                             bool time_kernel,
+                             std::vector<index_t> lengths)
+{
+    bool pass = true;
+
+    using ElementOp = ck::tensor_operation::element_wise::PassThrough;
+    using UnaryOp   = ck::tensor_operation::element_wise::UnarySquare;
+    using Scale     = ck::tensor_operation::element_wise::Scale;
+    float scale     = 2.f;
+
+    index_t N = lengths[0];
+    index_t C = lengths[1];
+    index_t H = lengths[2];
+    index_t W = lengths[3];
+
+    std::vector<ck::index_t> nchw = {N, C, H, W};
+    std::vector<ck::index_t> nhwc = {N, H, W, C};
+    Tensor<ADataType> a(nchw);
+    Tensor<BDataType> b(nhwc);
+    Tensor<BDataType> host_b(nhwc);
+
+    std::array<ck::index_t, 4> ab_lengths;
+
+    std::array<ck::index_t, 4> a_strides = {1,
+                                            static_cast<int>(nchw[0]),
+                                            static_cast<int>(nchw[0] * nchw[1]),
+                                            static_cast<int>(nchw[0] * nchw[1] * nchw[2])};
+
+    std::array<ck::index_t, 4> b_strides = {1,
+                                            static_cast<int>(nhwc[0] * nhwc[1] * nhwc[2]),
+                                            static_cast<int>(nhwc[0]),
+                                            static_cast<int>(nhwc[0] * nhwc[1])};
+    ck::ranges::copy(nchw, ab_lengths.begin());
+
+    std::cout << "A: " << a.mDesc << std::endl;
+    std::cout << "B: " << b.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1: a.GenerateTensorValue(GeneratorTensor_2<ADataType>{-1, 2}); break;
+    default: // a.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}
+        std::mt19937 gen(11939);
+        std::uniform_int_distribution<int> dis(0, 1);
+        auto i = 0;
+        for(std::size_t w = 0; w < a.mDesc.GetLengths()[3]; ++w)
+            for(std::size_t h = 0; h < a.mDesc.GetLengths()[2]; ++h)
+                for(std::size_t c = 0; c < a.mDesc.GetLengths()[1]; ++c)
+                    for(std::size_t n = 0; n < a.mDesc.GetLengths()[0]; ++n)
+                    {
+                        a.mData[(n * nchw[1] * nchw[2] * nchw[3]) + (c * nchw[2] * nchw[3]) +
+                                (h * nchw[3]) + w] = i;
+                        i                          = dis(gen);
+                    }
+    }
+
+    DeviceMem a_device_buf(sizeof(ADataType) * a.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b.mDesc.GetElementSpaceSize());
+
+    a_device_buf.ToDevice(a.mData.data());
+
+    std::array<const void*, 1> input = {a_device_buf.GetDeviceBuffer()};
+    std::array<void*, 1> output      = {b_device_buf.GetDeviceBuffer()};
+    using DeviceOp = ck::tensor_operation::device::DeviceElementwise<ck::Tuple<ADataType>,
+                                                                     ck::Tuple<BDataType>,
+                                                                     ElementOp,
+                                                                     UnaryOp,
+                                                                     Scale,
+                                                                     NumDim>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_instance_name;
+    float best_ave_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+    float best_tflops     = 0;
+
+    if(do_verification)
+    {
+        host_elementwise4D(host_b, a, ElementOp{}, UnaryOp{}, scale);
+    }
+
+    for(auto& op_ptr : op_ptrs)
+    {
+        auto argument_ptr = op_ptr->MakeArgumentPointer(ab_lengths,
+                                                        {a_strides},
+                                                        {b_strides},
+                                                        input,
+                                                        output,
+                                                        ElementOp{},
+                                                        UnaryOp{},
+                                                        Scale{scale});
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            b_device_buf.SetZero();
+
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+
+            if(do_verification)
+            {
+                b_device_buf.FromDevice(b.mData.data());
+
+                pass &= ck::utils::check_err(
+                    b.mData, host_b.mData, "Error: Incorrect results b", 1e-3, 1e-3);
+
+                if(do_log)
+                {
+                    LogRangeAsType<float>(std::cout << "a : ", a.mData, ",") << std::endl;
+                    LogRangeAsType<float>(std::cout << "b: ", b.mData, ",") << std::endl;
+                }
+            }
+
+            std::string op_name = op_ptr->GetTypeString();
+
+            float ave_time =
+                invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+            std::size_t flop = std::size_t(2) * nchw[0] * nchw[1] * nchw[2] * nchw[3];
+
+            std::size_t num_btype = sizeof(ADataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]) +
+                                    sizeof(BDataType) * (nchw[0] * nchw[1] * nchw[2] * nchw[3]);
+
+            float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                best_instance_name = op_name;
+                best_tflops        = tflops;
+                best_ave_time      = ave_time;
+                best_gb_per_sec    = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl;
+        }
+    }
+    if(time_kernel)
+    {
+        LogRange(std::cout << "length = ", lengths, ",") << ", ";
+        std::cout << "best perf = " << best_ave_time << " ms, " << best_gb_per_sec << " GB/s, "
+                  << best_instance_name << std::endl;
+    }
+
+    return true;
+}
+
+} // namespace ck